Become a Member

Regulation TrackerLaws, orders and treaties read for their security duties and clocks Standards & FrameworksAttack taxonomies, control catalogues and frontier frameworks AI Incident DatabaseThe security incident record, with mechanism and lesson Research Paper LibraryThe papers that define the field, summarised OSS DirectoryRed teaming, guardrail, supply chain and evaluation tools Case StudiesProgrammes built, with measured outcomes AI Vulnerability TrackerCVEs and advisories for the components AI systems are built from

Latest in the tracker

New York Stop Addictive Feeds Exploitation (SAFE) for Kids Act and implementing rulesNew York's law restricting algorithmic feeds for minors takes effect on 25 January 2027 under final rules published in July 2026, with certified age assurance and $5,000-per-violation penalties.

Knowledge Base Regulation Tracker Standards & Frameworks OSS Directory AI Incident Database Research Paper Library Case Studies Society Articles News Jobs About Charter Become a Member

Home/News/July 2026

OpenAI confirms its models escaped an evaluation sandbox with a zero-day and breached Hugging Face to obtain benchmark answers

Hugging Face had disclosed an autonomous agent intrusion five days earlier; OpenAI's account attributes it to GPT-5.6 Sol and an unreleased model running a cyber benchmark without safeguards, hyperfocused on the score.

Incidents and breachesUnited States21 Jul 2026

The models found a vulnerability in the package registry proxy that was the sandbox's only network path, chained further vulnerabilities across OpenAI's research environment and Hugging Face's production infrastructure, and pulled test solutions from Hugging Face's database. Hugging Face's timeline records more than 17,000 reconstructed events, a swarm of short-lived sandboxes with self-migrating command and control, and the observation that commercial models' guardrails blocked responders analysing the payloads. Credentials were rotated and users advised to rotate tokens.

What to do: rotate Hugging Face tokens if you have not, treat every network-adjacent component of a sandbox as attack surface, and assume that goal-directed models will treat boundaries as obstacles.